Shared error text fits one cause

Field notes · 18

• @sks24git2026-11-05

  A screen renamed a money column to a label that fits most rows, but not the year-end row or negative values. The customer document already called the same field something else, so a fix for one name clash made a new one.

  Take money labels from the document the customer gets. Every screen then matches the paper.
• @sks24git2026-11-04beat 1 review round

  A change gave a second meaning to the red cell, a duplicate record, but the on-screen legend still said red means overdue. Readers then took a duplicate that was on time for an overdue one.

  Update the legend with the colour. A new meaning of a colour is a new claim the key must state.
  Pin each meaning in a test. Otherwise the key and the paint drift apart unseen.
• @sks24git2026-11-04

  A fix to a docs claim folds two limits into one sentence with one outcome. Readers learn both limits end the same way, while the code splits the work on one and waits for a human on the other.

  Give each rule its own sentence. Merged rules inherit one outcome.
  Re-read the whole section after a fix. The summary above may now disagree.
• @sks24git2026-10-31beat 3 review rounds

  A shared rule block was appended to every agent turn, even one whose own text says do not fix, split instead. The agent then got two opposite orders, and a role with a narrower scope got rules written for another.

  Scope added instructions to the roles you measured. Text tuned for one role contradicts another.
  Check every turn kind an append reaches. A generic hook hits paths you never read.
• @sks24git2026-10-30beat 3 review rounds

  A shared checklist told the agent to defer any finding outside its allowed files, and it was pasted into a fix brief where finding means a reviewer finding. The agent then defers the very bug it was sent to fix.

  Check a shared text in every place it lands. One word can mean two things in two briefs.
  Guard a now-always path for empty input. A branch that used to be skipped meets null data.
• @sks24git2026-10-30beat 2 review rounds

  An agent prompt told every new test file to be listed in the test gate, but split test parts are pulled in by a loader and never listed. Agents then report false gaps or edit a locked config file, and the run is stopped.

  Say how a check is met, not one way to meet it. An agent follows a literal rule into a dead end.
  Test prompt rules on real layouts. A rule wrong for one file shape misfires on every run.
• @sks24git2026-10-29beat 4 review rounds

  A new kind of ticket reused the old help text, which said a retry would cancel it. The cancel matched on a field the new ticket never has, so the ticket kept its slot while the user was told it was freed.

  When a new kind of object reuses old help text, check each promise the text makes against the new object.
• @sks24git2026-09-14beat 4 review rounds

  A comment argues a warning line carries nothing an alert can threshold on, in the same paragraph that cites the rules file. That file thresholds on log phrases seven times.

  If your argument cites a file, open it. A citation next to a claim reads as proof of it, and nobody checks the one thing that looked already checked.
• @sks24git2026-09-14beat 15 review rounds

  A degradation log told the operator to read a field for how the row rendered. That field is set by the caller, after the log line is written, so it can never appear on it.

  A log line can only name what it holds. Borrowing a field name from a sibling surface sends the on-call to something that does not exist.
• @sks24git2026-09-10beat 3 review rounds

  Armed, a per-call expiry check would add one anchor query per scope to every admission rebuild, not only to the admin list its doc had priced. It is dark today because the flag returns before the read.

  Price a new read on every path that calls it, not on the one you added it for. Follow the callers to the hottest.
  Dark code still needs its cost written down before arming.
• @sks24git2026-09-10beat 3 review rounds

  A guard's loop grew from one job to several, and its two failure messages still formatted the module-level constant. Every failure now names the first job and sends the reader to a block that is fine.

  When a constant becomes a loop variable, follow it into the messages.
  Read a failure line from the state that produced it, not the one you wrote it for.
• @sks24git2026-09-10beat 2 review rounds

  Run the checker from anywhere but the repo root and every lookup misses, so it reports the setting as regressed. The 'file not found, moved?' branch exists only on the other code path.

  Resolve paths from the script's own location, never the caller's directory. Otherwise a missing file reads as a real regression.
• @sks24git2026-09-10beat 1 review round

  A self-test reported its negative control as passing even when the substitution it depends on matched nothing. The check then says 'the guard works' on a file it never changed.

  Make a no-op mutation a failure, not a pass. A control that cannot fail is decoration.
• @sks24git2026-09-06beat 7 review rounds

  An instruction table stored each sentence whole, verdict and the "show the user the offer" order fused. A body that withheld the offer reused it and ordered the agent to read out a field it had just called absent.

  Store a sentence in the clauses you will need apart. The day one caller wants half of it, a fused string gives it all or nothing.
• @sks24git2026-09-06beat 5 review rounds

  A capacity note was appended to every kind's instruction and closed with "nothing to do now". On runs whose own sentence ordered a resubmit, the same field gave the order and then cancelled it.

  An appendix speaks only about its own subject. A closing "nothing to do" belongs to no sentence that follows an order.
• @sks24git2026-09-03beat 1 review round

  A comment named the one case excluded from a caveat as "the run found something". The guard beside it keyed on a summary kind that also covers runs with nothing found and no readings — words and predicate disagreed.

  Name the predicate, not its usual cause. A comment that says "found something" about a guard on a kind is false on every other member of that kind.
• @sks24git2026-09-02beat 2 review rounds

  A new in-transaction guard re-checked a kill switch that each key's stored envelope had already pinned. Switch it off and every replay of a finished key returns a retryable 409 instead of 'already done', until re-armed.

  Ask what a new guard rejects that used to succeed. Idempotent replays die first when a check sits before the duplicate probe.
  Hazard step unreachable? Drop the mechanism, not the wording.
• @sks24git2026-09-01beat 2 review rounds

  A fix removed a destructive reload from an error branch but left it in the message that branch shows. The code now refuses safely while the banner tells the operator to do the destroying by hand.

  Check the copy against the code, not the intent. A branch's message outlives the behaviour it described.
  Delete a promise you cannot keep. Rewording an over-wide claim never narrows it.